Research on methods to differentiate coal and gangue using image processing and a support vector machine

ABSTRACT

This article proposes a method that can be used to improve the differentiation of coal and gangue via image processing and use of a support vector machine (SVM). Images of coal and gangue were converted to grayscale in this approach, the background was segmented, and the contrast was stretched. A basic eigenvalue was then determined based on the contrast between the grayscale mean and the gray-level co-occurrence matrix in each image. The biorthogonal wavelet was then used to expand coal and gangue images based on discrete wavelet transforms in two dimensions (2-D), while the supplementary eigenvalue is comprised of the mean variance of the wavelet coefficient at different scales. The eigenvalue of coal was then contrasted with each gangue eigenvalue, as well as the basic and the supplementary eigenvalue to construct a mathematical recognition model based on image processing and use of a SVM. At the same time, the penalty factor and kernel function coefficient of the mathematical model were optimized using K-fold cross validation. Experimental results indicate that the method proposed in this article can be used to recognize coal and gangue more effectively (at a rate up to 95.12%), compared to the conventional image processing recognition method.

KEYWORDS:

• Coal and gangue recognition
• image processing
• wavelet transform
• support vector machine

Highlights

1. Contrast in the gray-level co-occurrence matrix processed by histogram equalization can therefore very well represent the surface featured of coal and gangue.

2. A biorthogonal wavelet can be used to transform coal and gangue images and a method is proposed that the eigenvector is composed of mean of grayscale, contrast of gray-level co-occurrence matrix, and mean of variance of wavelet coefficients after wavelet transform.

3. Building a coal and gangue recognition model based on the use of a SVM optimizes the kernel function parameter as well as the penalty-factor. The recognition accuracy rate was as high as 95.12%.

Acknowledgment

This work is supported by the Yue Qi Young Scholar Project, China University of Mining & Technology, Beijing. We thank International Science Editing (http://www.internationalscienceediting.com) for editing this manuscript.

Additional information

Funding

This work was supported by the Yue Qi Young Scholar Project, China University of Mining & Technology, Beijing.